Modulation of aperiodic EEG activity provides sensitive index of cognitive state changes during working memory task
- University Psychiatric Clinic Ljubljana, Slovenia, Slovenia
- Faculty of Medicine, University of Ljubljana, Slovenia
- Department of Neurology, University Medical Centre Ljubljana, Slovenia
- Department of Psychology, Faculty of Arts, University of Ljubljana, Belgium
- Department of Neuroscience, Laboratory for Neuro- and Psychophysiology, KU Leuven, Belgium
Figures
Figure 1
Schematic representation of the n-back task.
The n-back working memory task was conducted in two distinct modalities: visuospatial and verbal. In the 2-back condition, participants had to identify whether the current stimulus matched the one presented two steps previously. In the visuospatial modality, the target was a spatial location, whereas in the verbal modality, the target was a letter. In the 0-back condition, participants’ task was to respond to a predefined target, with the type of target corresponding to the task modality. Target stimuli are highlighted in orange.
Figure 2
Schematic of the analysis.
The time domain data were first transformed into the time-frequency domain by convolution with superlets (Moca et al., 2021). Next, the periodic and aperiodic components of the power spectrum density were estimated for each time point using FOOOF (Fitting Oscillations and One-Over-F) algorithm (Donoghue et al., 2020b). The aperiodic component was characterised by the aperiodic slope (the negative counterpart of the exponent parameter) and the offset, which together describe the underlying broadband spectral shape.
Figure 3 with 4 supplements
Effect of baseline correction and FOOOF decomposition on time-frequency decomposition in the n-back task.
Illustration of the effect of baseline correction and FOOOF decomposition on electroencephalography (EEG) time-frequency analysis. (A) 'Raw’ total power from time-frequency decomposition is difficult to interpret due to the power scaling of EEG power spectra (here representing average data across subjects, channels, and conditions). In panel B, we applied baseline correction (decibel conversion) using a pre-stimulus interval of –0.5 to –0.2 ss for comparison. The baseline correction showed a significant decrease in alpha and beta power from 0 to 1 s and a concomitant increase in low-frequency power lasting up to 2 ss. The observed changes were very similar across different choices of baseline correction (Figure 3—figure supplement 2). However, it’s unclear from the baseline-corrected data whether the observed changes in the low-frequency range reflect periodic or aperiodic contributions. To disentangle these components, we decomposed the time-frequency signal into periodic and aperiodic contributions using spectral parameterisation. (C) The periodic component includes only the parameterised spectral peaks, reconstructed from Gaussian fits to the power spectrum (see Methods for details). (D) The aperiodic component reflects the 1/f-like background activity. This decomposition suggests that changes in power in the low-frequency (delta and theta) range may largely reflect changes in aperiodic activity. See Figure 3—figure supplement 1 for the corresponding figure with a logarithmic y-axis.
Figure 3—figure supplement 1
Effect of baseline correction and FOOOF decomposition on time-frequency decomposition in the n-back task.
Same as Figure 3 but with a logarithmic y-axis.
Figure 3—figure supplement 2
The comparison of different baseline corrections on time-frequency decomposition results in the n-back task.
We compared four types of baseline correction (in columns): (a) decibel conversion (), (b) relative change (), (c) normalised change (), (d) absolute change (). We also considered three baseline periods (in rows): (a) from –500 to –200 ms, (b) from –300 to 0 ms, (b) from –500–0 ms. The selection of baseline correction type and period had minimal impact on the outcomes. Among the methods, absolute change correction exhibited a slight reduction in deviations from baseline. Nevertheless, the overall results were qualitatively similar for all baseline corrections. The data presented represent the average across participants, conditions, and electrodes.
Figure 3—figure supplement 3
Effect of baseline correction and FOOOF decomposition on time-frequency decomposition (control dataset).
As in the main analysis (Figure 3), baseline correction (B) revealed changes in alpha, beta, and theta power after stimulus presentation. FOOOF decomposition showed that changes in alpha and beta were periodic (C), whereas changes in theta power reflected task-related changes in aperiodic activity (D) (see also Figure 8—figure supplements 7 and 8).
Figure 3—figure supplement 4
Effect of baseline correction and FOOOF decomposition on time-frequency decomposition (control dataset).
Same as Figure 3—figure supplement 3 but with a logarithmic y-axis.
Figure 4 with 6 supplements
Correlations between baseline-corrected time-frequency and FOOOF-decomposed EEG activity.
For each FOOOF parameter, we estimated its similarity to baseline-corrected time-frequency EEG activity at each channel-frequency-timepoint using a linear mixed model. We observed strong correlations between periodic activity and baseline-corrected time-frequency activity in the alpha and beta ranges. In contrast, the exponent and offset showed strong correlations in the low-frequency range, with the exponent also showing negative correlations in the gamma range. These results suggest that changes in aperiodic activity appear as low-frequency power in baseline-corrected time-frequency plots. Note that while is strictly non-negative, we assigned its sign based on the beta coefficient from the fixed effect in the model to facilitate interpretation. The values presented are averaged across all channels; see Figure 4—figure supplement 1 for topographical distributions. Condition-specific correlations are shown in Figure 4—figure supplements 2–4.
Figure 4—figure supplement 1
Topographies of correlations between baseline-corrected time-frequency and FOOOF-decomposed EEG activity.
See Figure 4 for a heatmap of average across all channels.
Figure 4—figure supplement 2
Correlations between baseline-corrected time-frequency activity and FOOOF-decomposed activity, separately for each condition, for the exponent.
See Figure 4 for correlations with conditions as the random effect.
Figure 4—figure supplement 3
Correlations between baseline-corrected time-frequency activity and FOOOF-decomposed activity, separately for each condition, for the offset.
Figure 4—figure supplement 4
Correlations between baseline-corrected time-frequency activity and FOOOF-decomposed activity, separately for each condition, for the periodic activity.
Figure 4—figure supplement 5
Correlations between baseline-corrected time-frequency and FOOOF-decomposed EEG activity (control dataset).
The results were very similar to those obtained with the other two datasets (Figure 4, Figure 4—figure supplement 1, Figure 4—figure supplement 6).
Figure 4—figure supplement 6
Correlations between baseline-corrected time-frequency and FOOOF-decomposed EEG activity (item-recognition task).
The results were very similar to those obtained with the other two datasets (Figure 4, Figure 4—figure supplement 1, Figure 4—figure supplement 5), but in the item-recognition task, we also observed smaller correlations with aperiodic parameters in the alpha range.
Figure 5 with 6 supplements
Phase-autocorrelation function and its correlations with FOOOF parameters.
Phase-autocorrelation function (pACF) (Myrov et al., 2024) is an alternative measure of oscillations that is sensitive to rhythmicity rather than the amplitude. Patterns of pACF (A) were largely consistent with periodic activity (B), particularly in the alpha range, supporting the notion that pACF and periodic activity reflect the same underlying processes. In contrast, the correlations with exponent and offset were low (only results for exponent are shown as results for exponent and offset were essentially the same). See Figure 5—figure supplement 1 and Figure 5—figure supplement 2 for detailed figures.
Figure 5—figure supplement 1
Changes in phase autocorrelation function (index of rhythmicity) as a function of time.
Figure 5—figure supplement 2
Correlations between phase-autocorrelation function and FOOOF parameters.
Figure 5—figure supplement 3
Changes in phase autocorrelation function (index of rhythmicity) as a function of time (control dataset).
As with the other two datasets, the patterns of phase-autocorrelation function (pACF) were broadly consistent with the periodic activity estimated by FOOOF (see also Figure 5—figure supplement 4).
Figure 5—figure supplement 4
Correlations between phase-autocorrelation function and FOOOF parameters (control dataset).
Figure 5—figure supplement 5
Changes in phase-autocorrelation function (pACF, index of rhythmicity) as a function of time (item-recognition task).
As with the other two datasets, the patterns of pACF were broadly consistent with the periodic activity estimated by FOOOF (see also Figure 5—figure supplement 6).
Figure 5—figure supplement 6
Correlations between phase-autocorrelation function (pACF) and FOOOF parameters (item-recognition task).
Figure 6 with 8 supplements
Changes in periodic activity as a function of time.
(A) The inspection of periodic activity revealed strong activity in the alpha and beta frequency bands, with a sharp decrease at around 0.5 s post-stimulus. The power was stronger in the 0-back condition, compared to the 2-back condition (see also Figure 7). (B) Early beta activity was most prominent at occipital and frontal channels. Note that the colour scale ranges in panel B differ between frequency ranges.
Figure 6—figure supplement 1
Power spectra of periodic activity on the Fz channel for each participant, averaged over time (between 0.5 and 2 s).
Inspection of the individual periodic power spectra showed that only two participants, SKP-1001 and SKP-2022, exhibited periodic activity in the theta band. Both axes of the spectra are displayed on a logarithmic scale.
Figure 6—figure supplement 2
Power of periodic activity 0.5 to 1 s post-stimulus, averaged across all channels and participants.
This visualisation facilitates the interpretation of the interaction between stimulus type and modality. It illustrates that the difference in power between target and non-target stimuli is more pronounced in the verbal task than in the visuospatial task. Error bars represent 95% Cousineau-Morey (within-subject) confidence intervals.
Figure 6—figure supplement 3
Changes in periodic (oscillatory) activity as a function of time.
Similar to Figure 6, but without subtracted event-related potentials (ERPs). The periodic activity remains very similar to that observed with the subtracted ERPs, with a notable increase in power at occipital and frontal electrodes from 0 to 0.5 s post-stimulus.
Figure 6—figure supplement 4
Power spectra of periodic activity on the E15 channel for each participant, averaged over time (control dataset).
Similar to our main analysis (Figure 6—figure supplement 1), theta periodic activity was reliably present in only one participant (1115), with another participant (1104) showing a peak at 7-8 Hz. Both axes are displayed on a logarithmic scale.
Figure 6—figure supplement 5
Changes in periodic (oscillatory) activity as a function of time (control dataset).
(A) Time course of periodic activity averaged across all channels. The periodic activity was most pronounced in the alpha band, with a strong decrease after stimulus onset. (B) Alpha and beta activity was most prominent in frontal channels.
Figure 6—figure supplement 6
Periodic activity on the item-recognition task.
Vertical lines represent: instruction presentation (0 s), stimuli presentation (0.4 s), end of encoding (0.8 s in load 2; 1.6 s in load 4), end of retention period (2.8 s). Prominent task-related modulations were observed in alpha and beta ranges, with a decrease during stimulus presentation, a relative increase during the retention phase and a decrease during probe presentation.
Figure 6—figure supplement 7
The power spectra of periodic activity on the Fz channel for each participant in the item-recognition task, averaged over time.
While the group average did not show pronounced theta activity (Figure 6—figure supplement 6, Figure 6—figure supplement 8), several participants showed periodic activity in the theta range (Seq2-s18, Seq2-s34, Seq2-s52, Seq3-s38, Seq3-s41, Seq3-s44, Seq3-s50). Furthermore, several peaks were observed between 7 and 8 Hz (Seq1-s14, Seq1-s23, Seq2-s37, Seq2-s40, Seq2-s55, Seq3-s28, Seq3-s59). Both axes are displayed on a logarithmic scale.
Figure 6—figure supplement 8
Group average power spectra of periodic activity in the item-recognition task, averaged over time.
As in Figure 6—figure supplement 6, only alpha and beta peaks were observed in the group average.
Figure 7 with 1 supplement
Results of the linear mixed model analysis of periodic activity for comparison between conditions.
Significant values are highlighted on the heatmap and marked with yellow circles on the topographies. Only factors with significant differences are shown. A p-value was interpreted as significant if it was significant in at least three channels or three time points (see Methods for details). Note that there are small discrepancies between significant values on heatmaps and topographies due to averaging across time or channels. A significant difference was observed between the 2-back and 0-back conditions, with a reduction in activity in the 2-back condition, particularly in the alpha and beta frequency bands, starting at 0.5 s post-stimulus. The differences between stimulus types were most evident from 0.3 to 1 s post-stimulus, with decreased activity for targets compared to non-targets across the whole scalp. Additionally, a smaller effect of the modality × stimulus type interaction was observed from 0.6 to 1 s post-stimulus (see also Figure 6—figure supplement 2 for detailed visualisation of the interaction). Associations with reaction times were significant in the beta band across central channels. These associations exhibited a positive correlation in the early phase (0–0.5 s) and a negative correlation in the later phase (1–1.5 s).
Figure 7—figure supplement 1
Results of linear mixed model on periodic activity for comparison between conditions (control dataset).
There was a significant effect of load with reduced alpha and beta power at higher loads.
Figure 8 with 10 supplements
Changes in aperiodic activity (exponent, interpreted as aperiodic slope) as a function of time.
(A) We averaged the time course of the aperiodic activity (exponent) over all channels and observed two components (peaks) of the aperiodic activity. The first component peaked around 0.3 s post-stimulus in the frontal channels. The second component peaked at 0.7 s, was stronger in parietal channels and differed between non-target and target conditions (see Figure 9). The time course of the offset parameter was comparable, although the separation between the frontal and parietal components was more pronounced (see Figure 8—figure supplement 1). The vertical lines in panel A represent mean reaction time modes. The shaded areas represent 95% Cousineau-Morey within-subjects confidence intervals, adjusted so that non-overlapping intervals correspond to statistically significant differences (Cousineau, 2017).
Figure 8—figure supplement 1
Temporal changes in aperiodic activity (offset).
Similar to Figure 8, but for the offset parameter. The vertical lines in panel A represent mean reaction time modes. The shaded areas represent 95% Cousineau-Morey within-subjects confidence intervals.
Figure 8—figure supplement 2
Grand average event-related potentials (ERPs) on midline electrodes.
ERPs have been low-pass filtered at a 40 Hz cut-off for visualisation.
Figure 8—figure supplement 3
Topographies of stimulus- and response-locked ERPs.
Figure 8—figure supplement 4
Correlations between ERPs and aperiodic activity.
Associations between stimulus-locked ERPs and aperiodic activity were estimated using a linear mixed model for each channel-time point. Correlations between ERPs and aperiodic activity peak right after stimulus onset but remain generally low. Notably, moderate correlations are observed between 0.5 and 1 s after stimulus onset on one channel, which we suspected were due to noise. To address this, we recalculated the model by capping values below the 0.1th percentile and above the 99.9th percentile (B, D). The results confirm that the associations between stimulus-locked ERPs and aperiodic parameters are very weak beyond 0.5 s.
Figure 8—figure supplement 5
Changes in aperiodic activity (slope or exponent) as a function of time.
Similar to Figure 8, but without the subtraction of ERPs. Aperiodic activity remains broadly similar to that observed with subtracted ERPs, with an increase in power at occipital electrodes and greater differentiation between frontal and parietal/occipital components. The vertical lines in panel A represent mean reaction time modes. The shaded areas represent 95% Cousineau-Morey within-subjects confidence intervals.
Figure 8—figure supplement 6
Changes in aperiodic activity (offset) as a function of time.
Similar to Figure 8—figure supplement 1, a frontal and parietal aperiodic component can also be observed when the ERPs are not subtracted. The vertical lines in panel A represent mean reaction time modes. The shaded areas represent 95% Cousineau-Morey within-subjects confidence intervals.
Figure 8—figure supplement 7
Changes in aperiodic activity (exponent) as a function of time (control dataset).
Similar to the main analysis (Figure 8), we observed two peaks, early frontal and late parietal components. The vertical lines in panel A represent mean reaction time modes. The shaded areas represent 95% Cousineau-Morey within-subjects confidence intervals.
Figure 8—figure supplement 8
Changes in aperiodic activity (offset) as a function of time (control dataset).
The vertical lines in panel A represent mean reaction time modes. The shaded areas represent 95% Cousineau-Morey within-subjects confidence intervals.
Figure 8—figure supplement 9
Aperiodic activity (exponent) in the item-recognition task.
As with periodic activity, we also observed task-related changes in aperiodic activity. Specifically, the aperiodic slope increased during instruction, stimulus presentation, and probe presentation compared to the retention period. The topographies were similar to those of the n-back task, with activity spanning frontal and parietal channels. However, the two components could not be distinguished (see also Figure 8—figure supplement 10).
Figure 8—figure supplement 10
Aperiodic activity (offset) in the item-recognition task.
The outcomes were comparable to those of exponent (Figure 8—figure supplement 9); however, the frontal and parietal/occipital components were discernible.
Figure 9 with 3 supplements
Results of the linear mixed model analysis of aperiodic exponent (interpreted as slope) for comparison between conditions.
Significant values are shown in blue on line plots and marked with yellow circles on topographies. Only factors with significant differences are shown. The only significant differences between conditions were observed between target and non-target conditions, where the exponent was higher for targets in the early phase (0–0.5 s post-stimulus) and for non-targets in the middle phase (0.5–1 s post-stimulus). There was also a small association with reaction times in occipital channels around 1.5 s post-stimulus. Results were similar for the offset parameter, with an additional effect of load and n-back × stimulus type interaction (Figure 9—figure supplement 1).
Figure 9—figure supplement 1
Results of the linear mixed model on aperiodic activity (offset parameter) of the comparison between conditions.
Significant values are shown in blue on line plots and marked with yellow circles on topographies. Results were similar to the exponent parameter (Figure 9), but there was also an effect of load at about 0.6 s post-stimulus.
Figure 9—figure supplement 2
Results of linear mixed model on aperiodic activity (exponent) for comparison between conditions (control dataset).
There was a significant effect of load, most evident immediately after stimulus presentation. There was also an association with reaction times, particularly after 1 s post-stimulus.
Figure 9—figure supplement 3
Results of linear mixed model on aperiodic activity (slope) for comparison between conditions (control dataset).
Figure 10 with 2 supplements
The effect of baseline correction and FOOOF decomposition on time-frequency decomposition in the item-recognition task.
Similar to the n-back task (Figure 3, Figure 3—figure supplement 3), a decrease in alpha power is observed following instruction and stimuli presentation (up to 0.8 s) and continues throughout the retention period, up to 2.8 s post-stimulus (B). This is followed by a decrease in alpha and beta power during probe presentation, which is likely indicative of a motor response. Simultaneously, there is an increase in low-frequency power, which is most pronounced during stimulus presentation (up to 0.8 s) and again after probe presentation (after 2.8 ss). The FOOOF decomposition indicates that a substantial portion of low-frequency activity could be attributed to the aperiodic component (C, D) (see also Figure 4—figure supplement 6). The data shown represent the group average over all conditions at electrode Fz, where low-frequency activity was most pronounced (see also Figure 6—figure supplement 8). Horizontal lines indicate the boundaries of the frequency ranges. See Figure 10—figure supplement 2 for the corresponding figure with a logarithmic y-axis.
Figure 10—figure supplement 1
Schematic representation of the item-recognition task.
In each trial, two or four target stimuli of a given condition were presented sequentially in each half of the visual field while the participant focused on one half. Participants held the relevant visual properties of the stimuli in working memory. This was followed by a probe in which two stimuli were presented, one in each half of the visual field, and participants responded whether the stimulus in the attended half matched one of the previous targets. Each trial began with a fixation cross in the centre of the screen on which participants maintained their gaze. Each trial began with instructions indicating the area of focus and the number of stimuli that would follow. After a blank interval, participants viewed the target stimuli, followed by a maintenance interval and a probe. During this interval, participants indicated whether the probe corresponded to one of the previous targets. Trials were counterbalanced to ensure that an equal number of presentations occurred in both visual fields for each of the conditions and working memory loads. The number of stimuli presented sequentially (i.e. working memory load) was randomised to prevent participants from making predictions.
Figure 10—figure supplement 2
The effect of baseline correction and FOOOF decomposition on time-frequency decomposition in the item-recognition task.
Same as Figure 10 but with a logarithmic y-axis.
Appendix 1—figure 1
Simulated power spectra in log-log space.
We added a single periodic component to each power spectrum, with the exponent fixed at 1.5 and the offset at 1.
Appendix 1—figure 2
Exponent estimates.
Red lines indicate ground truth. Panel A covers 1–30 Hz fits, while panel B covers 3–30 Hz fits. Exponent estimates were inflated at larger bandwidths and higher periodic power, especially when the periodic component’s central frequency was low. This suggests a mixing of periodic and aperiodic components.
Appendix 1—figure 3
Offset estimates.
As in Appendix 1—figure 2, estimates tended to be inflated when the periodic component was difficult to detect.
Appendix 1—figure 4
Bandwidth estimates.
Red lines show ground truth, while red numbers mark how often the periodic component was undetectable (out of 100 cases). Low frequency, low power, and the restricted 3–30 Hz fitting range often made periodic features hard to detect.
Appendix 1—figure 5
Central frequency estimates.
Central frequency estimates were inflated for large bandwidths and high power, particularly when fitting was restricted to ≥3 Hz (panel B). Variability of estimates also increased at low simulated power. The red numbers indicate how many times out of 100 the periodic component was undetectable.
Appendix 1—figure 6
Power estimates.
In contrast to bandwidth and central frequency estimates, power estimates remained largely accurate at low simulated power, regardless of bandwidth. However, at high simulated power and large bandwidths, power estimates were deflated. The red numbers indicate how many times out of 100 the periodic component was undetectable.
Appendix 1—figure 7
index of model fit.
values were very high in most cases, even when parameter estimates were largely inaccurate (see previous figures). The lowest values were observed for low bandwidths, low central frequencies, and high power.
Appendix 1—figure 8
Model errors.
Similar to , higher model errors were associated with low central frequencies and large powers, and the effect was more pronounced when only frequencies ≥3 Hz were fitted.
Appendix 2—figure 1
Fitting Oscillations and One-Over-F (FOOOF) goodness-of-fit measures.
Averaged across channels, subjects, and conditions. Goodness-of-fit was lowest around 0.5 s after stimulus onset, corresponding to the decrease in aperiodic and periodic activity.
Appendix 2—figure 2
Number of identified peaks in FOOOF models.
(A) Average number of peaks across all subjects, channels, and conditions. Light lines represent individual channels. (B) Histogram of the number of peaks in all models. On average, 1.9–2 peaks were identified per model, showing high consistency across models and indicating that the models did not overfit by detecting an excessive number of peaks.
Appendix 2—figure 3
FOOOF goodness-of-fit measures (control dataset).
Averaged across channels, subjects and conditions. Similar to our primary dataset (Appendix 2—figure 1), the decrease in goodness-of-fit around 0.5 s after stimulus onset coincided with the decrease in periodic activity.
Appendix 2—figure 4
FOOOF goodness-of-fit measures for the item-recognition task, averaged across channels.
was above 0.97 and model error was below 0.045 throughout the task. There was a small decrease in goodness-of-fit around 3 s after stimulus onset, coinciding with a decrease in periodic activity, similar to the other two datasets (Appendix 2—figures 1 and 3).
Author response image 1
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Modulation of aperiodic EEG activity provides sensitive index of cognitive state changes during working memory task
eLife 13:RP101071.
https://doi.org/10.7554/eLife.101071.4
